AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

Are Samsung or Hynix even doing 2-Hi stacks these days?

The only way I'd see a 2GB stack making sense is if HBCC is fully working. Does Kaby Lake G's Polaris Vega have HBCC?
 
ToTTenTranz said:
Are Samsung or Hynix even doing 2-Hi stacks these days?

The only way I'd see a 2GB stack making sense is if HBCC is fully working. Does Kaby Lake G's Polaris Vega have HBCC?
Click to expand...

Even if they were would such a low end card be worth the cost of HBM? Even if so it'd have to scale better than just half a Vega 64, like 25% better, to hit the same benchmarks the Rx 580 does. Makes the card seem odd, but that's the apparent leak.
 
ToTTenTranz said:
Are Samsung or Hynix even doing 2-Hi stacks these days?

The only way I'd see a 2GB stack making sense is if HBCC is fully working. Does Kaby Lake G's Polaris Vega have HBCC?
Click to expand...

Of course they are. It just depends on what the buyer wants from hynix or samsung. If someone wants 2-Hi, then they will mass produce it. They have it in their marketing material and production is easier than 4 or 8-high.

Frenetic Pony said:
Even if they were would such a low end card be worth the cost of HBM? Even if so it'd have to scale better than just half a Vega 64, like 25% better, to hit the same benchmarks the Rx 580 does. Makes the card seem odd, but that's the apparent leak.
Click to expand...

Worth the cost? If it's your only possibility to compete it's of course worth the cost. RX580 is nearly non-existent in notebooks, because it needs too much power. HBM will help Vega keep the power down and get a few design wins. In the design phase AMD probably hoped for cheaper HBM, but the high price leads now to just 2gb Ram.
 
Vega 10 vs. Vega 20 side-by-side:

JSro31I.jpg
TWZvLW8.jpg


The new chip is a lot smaller than Vega 10, maybe around half of that. Maybe 250-300mm^2?
A significantly smaller chip suggests there's really no big upgrade in execution units, as rumors have been pointing out to the same 64 CU count. With twice the memory channels they may have increased ROP count to 128, and there's accelerated dot8 instructions for AI.

The big question now is clocks and power consumption, which could hint at Navi's expected performance too.
 
ToTTenTranz said:
Vega 10 vs. Vega 20 side-by-side:

JSro31I.jpg
TWZvLW8.jpg


The new chip is a lot smaller than Vega 10, maybe around half of that. Maybe 250-300mm^2?
A significantly smaller chip suggests there's really no big upgrade in execution units, as rumors have been pointing out to the same 64 CU count. With twice the memory channels they may have increased ROP count to 128, and there's accelerated dot8 instructions for AI.

The big question now is clocks and power consumption, which could hint at Navi's expected performance too.
Click to expand...

It does not seem to me to be half of Vega 10, more like between 60 and 66% of it. But without a rigorous scale to measure it, its hard to know.

Regarding clocks and power consumption, my completely uneducated guess goes for 1.8Ghz without sweating it like Vega 10 did, at 180-200W. I would be shocked if they would get it down to 150W, even with the die shrink.
 
It wont be a gaming card anyway right ? Only pro / instinct stuff ? I think the good thing is 7mn seem ready for amd products... Damn, I hope Navi will deliver. I'm good with my watercooled Vega FE, but for the sake of competition, we need AMD... (or Intel with their new gpu division but they won't be there for few years I guess)
 
I tried to calculate die size using screenshot from the videostream. My first result was 360-380 mm². Using this (a bit better) image, the die size seems to be slightly smaller, about 355-370 mm². It seems, that the die is about 50 % bigger than hypothetical shrink of Vega 10.
 
No way that's 75% the size of a Vega 10.

Videocardz have made their own measurements and came up with 285-298mm^2, using this picture:

448F6p5.jpg



It puts Vega 20 at ~60% the size of Vega 10.
We should also take into consideration that measurements for the original Vega 10 also came out wrong because the black resin expands upon the original chip's area.
 
I count 133*88 pixels on each HBM2 for the Vega 10 (11704 pixels total) and 128*83 (10624 total) for the Vega 20.
Wouldn't say "nowhere near" with a 9.23% difference, but jpeg compression could be playing tricks to either side.
 
I don't have time to count pixels, but I did cut HBM dies from the Vega 20 picture in Photoshop and put them over the ones in Vega 10 and it looked almost 10% smaller in each direction.

I do call that nowhere near when making measurements.
 
This article (direct link) contains the best image (I believe), so I recalculated it once again. The range seems to be 343-352 mm², so it will be probably a bit smaller than I posted previously. But it isn't under 300 mm² for sure. The source image has quite good resolution and sharp edges, so error shouldn't be higher than a few percents.
 
Assuming the HBM2 dies are similar sized on Vega 10 and 20, I get Vega 20 to be 67% of Vega 10, or 325mm^2.

Also assuming ~40% performance increase, that'll give Vega 20 a bandwidth/flop ratio around Fiji, much better than Vega 10.

Cheers
 
I used this image from The Verge (I found it linked from here, the one in the Verge article goes to a lower resolution thumbnail) for Vega 20 and this image in the Videocardz analysis for Vega 10.

Code: 
                                  Dimensions:                 Area:          
                                  upper         lower         upper    lower    middle 
                                  bounds        bounds        bound    bound  

Vega 20 top left HBM2 stack       142    92     139    88     13,064   12,232   12,645
Vega 20 top right HBM2 stack      140    92     136    88     12,880   11,968   12,420
Vega 20 bottom left HBM2 stack    142    91     137    87     12,922   11,919   12,416
Vega 20 bottom right HBM2 stack   141    92     137    86     12,972   11,782   12,371
Vega 20 HBM2 stack average                                    12,960   11,975   12,463

Vega 10 top HBM2 stack             89   135      85   131     12,015   11,135   11,571
Vega 10 bottom HBM2 stack          90   135      84   130     12,150   10,920   11,528
Vega 10 HBM2 stack average                                    12,083   11,028   11,549

Vega 20 package                   280   176     274   169     49,280   46,306   47,783
Vega 10 package                   228   300     224   296     68,400   66,304   67,348
All numbers in the above table are in pixels. The package boundaries in the pictures are not sharp as they have color gradients. The upper and lower bounds are the dimensions of rectangles that cover almost all and almost none of the boundary gradients respectively. The numbers in the "middle" column are found by taking the average of the upper and lower bounds for each dimension and multiplying them together.

Assuming that the HBM2 stacks are the same size across Vega 10 and Vega 20, that gives area scale factors of
12,960/11,028 = 1.18 (maximum), 11,975/12,083 = 0.99 (minimum), and 12,463/11.549 = 1.08 (middle).​

The Vega 10 die size is 484 mm^2, so the Vega 20 die size can be estimated as
49,280/66,304 · 1.18 · 484 mm^2 = 423 mm^2 (maximum),
46,306/68,400 · 0.99 · 484 mm^2 = 325 mm^2 (minimum),
47,783/67,348 · 1.08 · 484 mm^2 = 371 mm^2 (middle).​

So my Vega 20 die size estimate is ~371 mm^2.
 
Rootax said:
It wont be a gaming card anyway right ? Only pro / instinct stuff ?
Click to expand...

I think that not announcing a gaming SKU is probably just a way of hedging their bets. For example, if Nvidia releases their new generation and the way things shape up are that AMD can't produce a Vega 20 consumer product with better price/perf than Nvidia without selling for a loss then AMD could end up with some significant egg on their face if they commit ahead of time to a gaming SKU. This way they're free to release one or not depending on what the market dictates.
 
You must log in or register to reply here.
Back
Top